DeepMind swarm of 100 agents split into cheaters and whistleblowers after finding a grading exploit

Google DeepMind researchers ran 100 agents, all based on Gemini 3.1 Pro, in a simulated scientific conference and tasked them with solving 71 mathematical conjectures formalized in the Lean proof language. The agents shared base weights and core prompts but received randomized domain personas and minor specializations. They could communicate through a public forum, direct messages, and a shared knowledge library. Each agent's system prompt warned that proofs must be mathematically genuine and that bypassing verification would result in zero credit. The verification itself was shallow. It checked whether code looked formally correct and compiled, but not whether the proof was valid. After an agent called prover-theta found the exploit, fake solutions for the remaining 34 problems propagated through the shared library within 27 minutes. The swarm then fractured. According to the paper, 9 percent of agents exploited the flaw, 5 percent converted to cheating after seeing it, 24 percent acted as whistleblowers, and 62 percent kept solving problems unaware. Whistleblowers audited proofs, messaged peers, boycotted the task, and filed complaints without any human prompting, but had no tools to delete fakes or sanction others. The preprint, titled A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms, is arXiv:2609.04170, submitted on 3 September 2026. Tbreak found no matching DeepMind blog post.